Skip to main content
Internet Archive's 25th Anniversary Logo

Full text of "Machine learning of fault characteristics from rocket engine simulation data"

See other formats

N96- 70676 


•Min Ke 
Moonis All 

Center for Advanced Space Propulsion 

The University of Tennessee Space Institute 

Tullahoma, TN 37388 


Transformation of data into knowledge through 
conceptual induction has been the focus of our re- 
search described in this paper. We have developed 
a Machine Learning System (MLS) to analyze the 
rocket engine simulation data. MLS can provide to 
its users fault analysis, characteristics, and concep- 
tual descriptions of faults, and the relationships of 
attributes and sensors. All the results are critically 
important in identifying faults. 

I. Introduction 

An important component of intelligent diagnos- 
tic systems is the knowledge which human experts 
employ in analyzing and diagnosing faults. However, 
this knowledge employed is very limited in the sense 
that it is based on a very small number of observed 
situations. Since human experts have not seen all pos- 
sible instances of all faults, they cannot describe fault 
characteristics sufficiently well to make diagnostic de- 
cisions. We have developed a Machine Learning Sys- 
tem (MLS) for analyzing the SSME simulator data 
to generate characteristics about engine faults. In 
MLS inductive heuristics and domain knowledge are 
employed to guide the inductive process. With two 
phases of abstractions as well as a knowledge man- 
agement system, MLS can be applied to a wide spec- 
trum of domain tasks. MLS has been tested with 
SSME simulator data. MLS consists of two levels 
of abstractions. Section II presents the general al- 
gorithm, Section III describes the basic abstraction, 
Section IV describes the advanced abstraction, and 
Section V discusses the tests and results. 

II. The General Algorithm 

MLS consists of two levels of abstraction: the 
basic abstraction which generates the characteristic 
descriptions, discriminant descriptions and aggrega- 
tional descriptions for a concept such as an engine 
fault; and the advanced abstraction which groups sim- 
ilar concepts into a concept hierarchy to form a higher 

This research was supported by NASA Grant Nos. 
NAGW-1195, and NAG-1-513 and Rocketdyne Con- 
tract No. R04QBZ90-032709. 

of concepts. MLS incorporates concept instance data 

In MLS the paradigm for inductive inference can 
be formulated as follows: 

Before induction we have: 

• A hierarchy of nodes with leave nodes represent- 
ing basic concepts, internal nodes representing 
clusters, and descriptions of concepts at each 

• Knowledge bases which include knowledge about 
attributes, components, relationships and basic 
concepts, and also include deductive rules, gen- 
eralization rules, transformation rules and aggre- 
gation rules; 

• A new instance description of a basic concept 
represented in MLS's representation language. 

After induction we get: 
e An extended hierarchy of concepts such that ei- 
ther a new basic concept node is added or an old 
basic concept's descriptions are modified to cover 
the new instance; the structure of the hierarchy 
and the cluster node's descriptions are modified 
to incorporate the new instance. 

Let C be a basic concept, and let e be a new in- 
stance of C. A raw description is a description of raw 
data in our representation language. In the knowl- 
edge base, appropriate transformation rules, aggrega- 
tion rules, deductive rules, generalization rules and 
inductive heuristics are provided. It is assumed that 
e is a set of simple expressions (i.e., atoms). The fol- 
lowing is the outline of the general algorithm of MLS: 

1) Read in e, raw description of an instance of 
concept C. 

2) Apply simple transformation and attribute- 
level aggregation on e, generating result el. 

3) Apply deductive transformation and 
component-level aggregation on el, generating result 
e2. New attributes and relationships are generated in 
this step. 

4) Generalize concept C's characteristic descrip- 
tions to cover e2. 


5) Generalize concept C's discriminant descrip- 
tions and specialize (or eliminate) discriminant de- 
scriptions of other concepts. 

6) Modify the concept-level aggregation descrip- 

7) Modify the generalization hierarchy above the 
basic concept level. The most frequent operation is to 
modify (adding, generalizing, or changing the weight 
of) the descriptions of the concept class on the higher 
levels. Other operations include creating a new class 
and deleting an unqualified class. 

The algorithm described here is only one process 
of a single instance. As an incremental method, the 
above algorithm can be repeatedly applied to many 
incoming instances. 

III. Methods and Algorithms for the Basic Abstrac- 

MLS is an incremental learning system, so it is 
efficient to incorporate new instances. MLS has a ca- 
pability of rich logic representation. Multiple-valued 
nominal attributes and inexact value matching are 
MLS's capabilities not shared by most other systems. 

Each time an instance of a specified concept is 
incorporated into the concept hierarchy, the object of 
the concept and the object of the instance are sent to 
the procedure that incrementally modifies the char- 
acteristic description of the concept. The algorithm 
for construction of characteristic descriptions can be 
described as follows: 

1. Check whether the basic concept is a new con- 
cept. If it is a new concept, the instance descrip- 
tion is taken as the characteristic description of 
the concept and the concept hierarchy should be 
extended to incorporate the new basic concept. 
If it is not a new concept, then perform the fol- 
lowing steps. 

2. For each expression in the instance object, try 
to find the matching expression in the concept 

If not found, put the expression into the Al- 
ternative List AL. 

If found, generalize the two matching ex- 
pressions, use the resultant expression as one 
expression of the characteristic descriptions, 
increase the count of instances that imply 
this expression, and compute the worth of 
the new expression. 

3. Apply the Add- Alternative Rule to the un- 
matched expressions of the concept characteristic 
descriptions and expressions in AL. The resul- 
tant expressions are put into the concept char- 
acteristic description. Also some unqualified ex- 
pressions are eliminated from the characteristic 
description of the basic concept. 

4. Call the bottom-up modification procedure in 
the advanced abstraction. Use only confident fea- 
tures of the concept to modify the concept hier- 
archy. If the consistency factor of an expression 
is greater than the consistency factor threshold, 
then the expression is considered a confident fea- 
ture. Since all characteristic features are com- 
plete (completeness factor is 1.0), only the con- 
sistency factor needs to be considered. 

The Algorithm for Constructing Discriminant 
Descriptions is given below: 

The inputs to this algorithm are an object of con- 
cept CI and an object of Cl's instance. 

For each expression EXP in the instance descrip- 
tion perform the following steps: 

1. Obtain all matching expressions from the unique- 
ness table UT. 

2. If there is no matching expression from UT, then 
EXP is a unique feature. 

Add EXP to the discriminant description of CI. 

Add an entry of EXP in UT. Exit. 

3. If there are matching expressions, then try to find 
inconsistency. Let ML be the list of matching 
expression entries. 

For each entry in ML, check for inconsistency 

(Loop A) 
1) If the expression EXP is a discriminant fea- 
ture of another concept C2, then an incon- 
sistency situation is found. 

Delete EXP from the discriminant description of 


Set the entry uniqueness flag off. 

Exit Loop A. 

2) If EXP is covered by a discriminant feature 
of another concept C3, then an inconsistency 
situation is found. 

Specialize the discriminant feature of C3. 

Modify the entry in UT. 

Exit Loop A. 

3) If EXP is covered by an inconsistent feature 
(an inconsistent feature is a feature that is 
already identified as nondiscriminant), then 
EXP is inconsistent. Exit Loop A. 

4. If EXP is not found to be inconsistent in step 3, 
then check for partial inconsistency. 

For each entry in ML, if the expression in 
the entry is partially covered by EXP, then 
specialize both expressions; 


modify the discriminant descriptions of 
the concept specified in the entry; 
modify the entry. 

5. If EXP is matching a discriminant feature EXP 1 
of CI (in which case EXP is said to be compati- 
ble), then 

perform the least generalization on EXP and 

modify Cl's discriminant description 

modify the entry in UT. 

6. If EXP is neither inconsistent nor compatible, 

add EXP to Cl's discriminant description; 
create an entry for EXP in UT. 

IV. Methods and Algorithms of the Advanced Ab- 

The advanced abstraction is an integration of two 
incremental processes: 1) modification of the clus- 
ter hierarchy, and 2) modification of cluster descrip- 
tions. The results of the advanced abstraction are a 
clustering of basic concepts and conceptual descrip- 
tions of the clusters. The combination of incremental 
processes with an expressive logic representation lan- 
guage makes MLS a unique system in the conceptual 
cluster area. In this section, we illustrate how the 
matching factor is computed; describe how to mea- 
sure the quality of clustering; discuss the operations 
on the concept hierarchy as well as the algorithms 
of hierarchy extension and modification; discuss the 
cluster parameters; and analyze the time cost of the 
advanced abstraction. 

Features in a cluster description will be con- 
stantly modified during the incremental inductive 
process. A feature has a confidence count which may 
be increased or decreased. The confidence count of a 
feature determines whether the feature is confident or 
not. Only the set of confident features is used as the 
description of a cluster. 

When the first instance of a new basic concept is 
incorporated into the concept hierarchy, a top-down 
extension process is performed. The procedure to per- 
form top-down extension of the concept hierarchy is 
a recursive one. The first call uses the root node as 
one of the parameters. The root is a universal clus- 
ter node which incorporates all instances input to the 
inductive system. The description of the root is a 
generalization of all the incorporated instances. 

A description of the procedure, nameed extend- 
hierarchy, is given below where curreninode is the cur- 
rent cluster, and newconcept is the new basic concept 
to be incorporated: 

extend-hierarchy (curreninode newconcept) 

(1) Modify the current cluster: 

a. Increase the size (by 1) of the current clus- 

b. For each expression E in the cluster descrip- 
tion perform the following steps: 

i. If there is no matching expression of 
E in newconcept's confident features, then see 
whether the expression confidence count of E is 
still greater than the feature retaining thresh- 
old FR-TH. Delete E if the expression confidence 
count of E is less than FR-TH. Retain E if the 
expression confidence count of E is not less than 

ii. If there is a confident expression El in 
newconcept that matches E and El is covered by 
E, then increase the expression confidence count 
of E by 1, and compute the worth of the expres- 

iii. If the confident feature El is not cov- 
ered by E, but the value matching factor of 
El and E is greater than the feature matching 
threshold FM-TH, then generalize E to incorpo- 
rate El and increase the expression confidence 
count of E by 1, and compute the worth of the 

iv. If El is not covered by E and value 
matching factor of El and E is less than FM-TH 
but greater than the feature conflict threshold 
FC-TH, then do nothing. If the value match- 
ing factor is less than FC-TH, then decrease the 
expression confidence count of E by 1, and see 
whether the expression confidence count of E is 
less than FR-TH. Delete E if the count is less 
than FR-TH. Retain E if the count is not less 
than FR-TH. 

v. Add all unmatched expressions of netu- 
concept's confident features to curreninode's de- 

(2) In the current cluster find each subcluster S that 
is close to newconcept and whose description does 
not violate constraints of neu/concepi. Closeness 
is measured by the matching factor between S 
and newconcept. A closeness threshold C-TH is 
used to determine whether a subcluster is close 
enough to newconcept. 

(3) Find each basic concept directly under the cur- 
rent cluster that is close enough to newconcpt. 

(4) If there are no close subclusters and basic con- 
cepts, then put neu/concepi directly under the 
current cluster. 

(5) Find the best object from those close subclus- 
ters and basic concepts according to the quality 
measure of clustering. Here the quality measure 
is modified to include basic concepts. Since a 
basic concept does not have a subcluster or sub- 
concept, the cluster matching factor is replaced 
by the matching factor of the basic concept and 


(6) If the best node is a basic concept, then com- 
bine the best node with newconcept to form a 
new subcluster of the current cluster: Create a 
new cluster. Generalize expressions of the de- 
scriptions of the two concepts. Two expressions 
are generalized only if their value matching fac- 
tor is greater than FM-TH (0.7). The new ex- 
pression's confidence count is the sum of the two 
counts of the generalized expressions. The new 
expression's worth is the average worth of the 
two generalized expressions. 

(7) If the best node is a subcluster, then recursively 

extend-hierarchy ( bestnode newconcept) 

When incorporating a new instance to an existing 
basic concept, the description of the basic concept 
may be generalized to cover the new instance. This 
modification may cause further modifications on the 
predecessors of the basic concept. 

For each expression EXP in the instance's de- 
scription, if it is not covered by a matching expres- 
sion of the basic concept, the following procedure is 

Let P be the parent cluster of the basic concept 
C, OLDEXP be the expression in C that matches 
EXP, NEWEXP be EXP or the generalization of EXP 
and OLDEXP. 

modify-hierarchy( C, OLDEXP, NEWEXP) 

Get C's parent P. 

Try to find EXP1, the expression in P that 
matches OLDEXP. 

If not found (EXP1 is empty), then add NEW- 
EXP to P's description. 

If EXP1 is not empty, NEWEXP is EXP, OLD- 
EXP covers NEWEXP, and the value matching factor 
of OLDEXP and EXP1 is greater than FM-TH, then 
increase the confidence count of EXP1 (Now NEW- 
EXP is taken as the supporting feature of EXP1). 

If EXP1 is not empty, NEWEXP covers OLD- 
EXP, and the value matching factor of NEWEXP and 
EXP1 is greater than FM-TH, then generalize EXP1 
to incorporate NEWEXP and increase the confidence 
count of EXP1. 

If EXP1 is not empty, NEWEXP covers OLD- 
EXP, and the value matching factor of EXP1 and 
NEWEXP is less than FC-TH, then NEWEXP is con- 
sidered to be contradict to EXP1. 

- The confidence count IC of EXP1 is decreased 
by 1. 

- Test IC to see whether it is less than FR-TH. 

- If IC is less than FR-TH, EXP1 is deleted from 
P's description; check the number of expressions 
in P's description; if the number is less than the 
cluster retaining threshold CR-Th, then delete P 
from the concept hierarchy and reassign all P's 

leaves and subclusters under P's parent (note: 

the root can never be deleted). 

If P has a parent, then recursively call: 

modify-hierarchy( P, OLDEXP, NEWEXP). 

After modifying the concept hierarchy, try to 
combine C with one of its sibling concept nodes and 
create a new cluster, since after the modifying of the 
concept description, the concept may become close to 
another basic concept under the same cluster node. 

V. Analysis and Results 

Engine test analysis is one of several application 
areas of inductive learning, where conceptual descrip- 
tions about different faults can be automatically gen- 
erated from a large number of fault instances, and 
similar faults can be classified into clusters. The in- 
ductive results, including high level characteristic de- 
scriptions and discriminant descriptions of faults and 
the concept hierarchy with descriptions of clusters, 
can be used to aid the fault test analysis and be used 
for fault diagnosis. 

The Space Shuttle Main Engine (SSME) is one 
of the most complex reusable liquid-fuel (oxygen and 
hydrogen) rocket engines. Each time a test on SSME 
is performed, a huge amount of data is collected from 
many sensors. Many highly-trained engineers are re- 
quired to perform a thorough investigation of the 
tests. Two difficulties are presented for the improve- 
ment of test analysis: (1) As more tests are performed 
and more thorough investigations are required, more 
experienced engineers are needed; (2) more senior 
staff with many year's experience are leaving. To 
overcome these difficulties, a computer conceptual in- 
duction technique is used to aid the engineers in an- 
alyzing the test data. In addition to its efficiency 
in forming concepts and generating concept descrip- 
tions, a computer inductive system can accumulate 
knowledge from both the data of many tests and the 
expertise of the engineering staff. 

After SSME simulator data is input into MLS, 
a concept hierarchy is built by the inductive system. 
The concept nodes on the hierarchy represent various 
engine faults. The cluster nodes on the hierarchy rep- 
resent higher level concepts, each of which describes 
a group of similar engine faults. Descriptions of en- 
gine faults or fault groups are also generated by the 
inductive system and stored in the nodes on the hier- 
archy. Features about any attributes, sensors or rela- 
tionships can be easily accessed by a user. The exper- 
tise of the engineering staff can be incorporated into 
the system as concept constraints, deductive rules and 
various inductive biases. 

SSME simulator generates the raw data about 
sensors for a fault. The raw data is simply a list of 
time-value pairs of each sensor. All values are in real 


number form. Usually this kind of data is used to 
plot diagrams (Figure 5.1) for representing sensor be- 
havior, and then human experts analyze the diagrams 
to find the characteristics of each fault. This human 
analysis process is usually time-consuming when the 
number of diagrams is large, and is complicated when 
the features of faults involve the relationships between 
sensors. Since a human expert describes the features 
of a fault by a set of attributes (which are shown in 
Table 5.1 and Table 5.3), MLS will automatically gen- 
erate those human-oriented descriptions from the raw 

Before sending the raw data to the induction sys- 
tem, preprocessing is performed which smooths the 
curves of sensors, divides the curves of sensors into 
segments and denotes each segment as an event. The 
collection of sensor descriptions constitutes the de- 
scription of an instance of an engine fault which, in 
turn, is used for the induction process. MLS empha- 
sizes a significant event for each sensor, since most 
characteristic features of a fault exist in the signifi- 
cant event. The preprocessor extracts basic attribute 
features from the raw data (as shown in Table 5.1) 
and deduction process generates more attribute fea- 
tures to describe a fault. The derived attributes are 
shown in Table 5.3. 

Although MLS can perform induction on a do- 
main without all the related domain knowledge, do- 
main knowledge makes the induction more time- 
efficient and produces better results. A discussion of 
the domain knowledge needed in the SSME applica- 
tion area is given below. 

A basic concept is an abstraction of a class of 
real world entities which share common features. A 
real world entity is a thing (such as an animal or a 
computer) or a situation (such as a disease, a ma- 
chine fault, or a state of a process). In the SSME 
fault test analysis, each type of engine fault is a basic 
concept. MLS assumes that every real world entity 
belongs to only one basic concept. In the SSME do- 
main we assume each fault instance represents only 
one fault. MLS takes in the instances of the basic 
concepts to incrementally generalize the descriptions 
about the basic concepts and to classify them into 
clusters. The basic concepts should be the main focus 
of an application domain if the purpose is to find the 
features of the basic concepts. In the SSME fault test 
analysis domain, the purpose is to find features for 
each fault and possible classification of faults. That 
is why engine faults are taken as basic concepts. In 
some application areas where the purpose is cluster- 
ing, we can use basic concepts to represent every real 
world entity. In this case, MLS does not perform the 
generalization in the basic abstraction; the main task 
performed is in advanced abstraction. 

The matching threshold for grouping basic con- 
cepts into classes is related to the number of lev- 
els in the hierarchy. In the SSME domain, faults 
can be grouped into several classes such as injec- 
tor faults, control faults, duct faults, manifold faults, 
valve faults, high pressure oxidizer turbopump faults, 
and high pressure fuel turbopump faults. The possi- 
ble number of levels of interesting high level concepts 
is one or two. By this kind of domain knowledge 
and purpose of clustering we choose the maximal level 
number to be three in MLS. 

A structural concept has components whose fea- 
tures and relationships collectively constitute the de- 
scription of a concept. For example, in a block world 
domain each block can be a component of a basic con- 
cept. In a cancerous cell analysis domain cell bodies 
are components of the basic concepts - cells. In a 
rocket engine fault analysis domain, an engine fault is 
described by features of temperature, pressure, flow, 
speed, etc. These parameters are measured by many 
sensors. Therefore, sensors are the components. 

In a multi-concept inductive system components 
usually have different relevancies with different con- 
cepts. For example, a sensor may have distinct fea- 
tures for an engine fault and show nothing about 
other faults. A large number of sensors exist in the 
rocket engine, but only a few of them are related to 
a specific fault. To pay equal attention to all sensors 
for every fault is inefficient. Therefore, a component 
in MLS can be assigned different relevancies for dif- 
ferent concepts. In building MLS the assignment of 
sensor-fault relevancies depends on domain knowledge 
such as functional relationships and structural rela- 
tionships of the engine parts as well as locations of 
sensors and faults. Examples of component-concept 
relevancies in MLS are shown in Table 5.2. In this 
table, slO, s22, etc. are sensors; CCV, MOV, MFV, 
OPOV, and FPOV are five types of engine faults. In 
MLS sensors can be in one of three types with re- 
spect to each engine fault: critical sensors — which 
show strong evidence of and are closely relate to the 
fault; irrelevant sensors — which do not show any 
changes when the fault occurs; and unspecified sen- 
sors — which may show some change with the fault 
and whose relationship to the fault is unknown. As 
indicated in Table 5.2, critical sensors are assigned 
a high relevancy value (10.0); irrelevant sensors are 
given a low relevancy value (2.0); and unspecified sen- 
sors are given a value between those two values (7.0 is 
assigned to a temperature sensor and 8.0 is assigned 
to a pressure sensor). From the domain knowledge we 
know that pressure sensors are usually more impor- 
tant in identifying a fault than temperature sensors. 
Thus, pressure sensors are given higher relevancy val- 
ues than temperature sensor. 

Relationships of component features play an 


important part in the description of concepts. Com- 
ponents may have positional relationships, temporal 
relationships, or some relationships governed by do- 
main theory. MLS. allow a system developer to in- 
dicate interesting component pairs. Then, deductive 
rules, which derive component-relationship descrip- 
tions from component features, are automatically gen- 
erated. Examples of these kinds of rules in MLS 
are shown in Figure 5.2. Known relationships be- 
tween component features can be used as concept con- 

Attributes are the basic vocabulary to describe 
basic concepts. After the identification of basic con- 
cepts the system developer needs to find out what at- 
tributes should be used in modelling the domain prob- 
lem. The choice of attributes is based on the avail- 
ability and utility. Availability tells what attributes 
can be abstracted directly from the input raw data. 
Usually too many attributes can be abstracted from 
the raw data, but only a small portion is relevant 
and useful to the concept description. The utility of 
an attribute, which tells what attributes are usually 
used to describe a basic concept, is determined by ex- 
perts with domain knowledge. Attributes determined 
by domain experts may not be available, so the con- 
structive rules should be formulated so as to derive the 
unavailable attributes from the available attributes. 

In MLS attributes are of different importance in 
describing a concept. Domain knowledge can be used 
to assign different worth values to different attributes. 
For example, in the SSME domain the attribute di- 
rection of change is more important than the at- 
tribute rate of change because a sensor's direction 
of change is usually the same for the same fault while 
the rate of change can be different for different sever- 
ities and durations of the fault. In many cases, rela- 
tionships between attributes are more important than 
individual attributes. MLS represents every attribute 
by an object and supports the knowledge acquisition 
facilities to help the system developer to define at- 
tributes. The "curve-pattern" attribute has a hier- 
archical domain which is represented by a list repre- 
sentation of tree. The slot "correspondence" is the 
transformation rule which transforms the input val- 
ues into the symbolic values (for example, the value 
'rf ' stands for a two-event curve pattern with the first 
event as 'rising' and the second event as 'falling'). 

Inductive rules include generalization rules and 
transformation rules. Generalization rules are sup- 
plied by the inductive system. Transformation rules 
are domain related. Domain knowledge is needed to 
determine how to divide a real number value-domain 
into categories, and what symbol represents a value 
range. For example, the value of attribute tempera- 
ture can be categorized into { high, very-high, 
medium, low, very-los } . For different domains 

the categories may cover different value ranges. There 
are no universal rules of transformation; the only cri- 
teria are that categories of values should correspond 
to the categories of concept instances, and that sym- 
bols need to reflect the value ranges in real world ap- 

Since useful attributes may not be available di- 
rectly from the raw data, deductive rules are used 
to derive them by applying various domain knowl- 
edge such as domain theory, physical laws, opera- 
tional principles and domain experience. From do- 
main knowledge in the flight engine test, we know 
that the attributes "starting time", "ending time", 
"changing rate" , and "magnitude" have little value to 
characterize an engine fault, because they all change 
with the severity or duration of an engine fault. Dif- 
ferent faults may have the same changing rate, and 
faults of the same type may have different changing 
rates. We found some relationships have more impor- 
tance in characterizing engine faults. For example, a 
temperature sensor and a pressure sensor at the same 
location of the engine have certain relationships for a 
specific fault. New attributes STAKT-TIME-DIFF 
(difference of the starting time) and END-TIME- 
DIFF (difference of the ending time) are used to repre- 
sent the temporal relationships; RATE-RATIO (ratio 
of the changing rate) and MAGNITUDE-RATIO (ra- 
tio of the magnitude) are used to represent the quan- 
titative relationships. The derived attributes in MLS 
are shown in Table 5.3. Logical relationships, like 
-the concurrency of changing trend, is represented by 
the logical connective AND. In the domain of SSME 
test analysis, human experts recognize certain fea- 
tures, and relationships for different faults. This kind 
of knowledge can be used as concept constraints. For 
example, (assuming an open loop situation) an in- 
creasing pressure of a valve inlet will cause an in- 
creasing pressure of the valve outlet. This rule is ap- 
plicable to all valve blockage faults. In MLS this rule 
is expressed as: 

IF (direction (s23) = ?x) 

THEN (direction (s27) = ?x) 

IF (direction (s23) = ?x) 

THEN (direction (s28) = ?x) 
where s23 is the pressure at the outlet of the high 
pressure oxidizer pump booster which is also the inlet 
to FPOV (fuel preburner oxidizer valve) and OPOV 
(oxidizer preburner oxidizer valve), s27 is the pressure 
of the outlet of OPOV, and s28 is the pressure of the 
outlet of FPOV. Examples of the deductive rules in 
MLS are shown in Figure 5.3. 

The SSME simulation data of 61 instances about 
five valve-blockage faults is used to run the indu- 
tive system. Attributes have different worth values 
and sensors have different relevancy values based on 
whether a sensor is a critical sensor to a fault, an 


irrelevant sensor or an unspecified sensor. 

The concept hierarchy of the induction is shown 
in Figure 5.4. We can see that the induction gives 
quite good clustering. There are no faults grouped 
with different types of faults on the second level. On 
the first level of the hierarchy, the clusters show strong 
regularity. Clusterl04 corresponds to the MFV block- 
age fault, cluster81 corresponds to the CCV block- 
age fault, cluster78 corresponds to the MOV block- 
age fault, and cluster73 corresponds to the OPV and 
FPOV blockage faults. In the hierarchy we can see 
that the OPV and FPOV faults have similar sensor 
behavior. Examples of the MSL output are shown 
in Table 5.4. and Table 5.5. In Table 5.4. a part 
of the characteristic description of Main Fuel Valve 
(MFV) blockage fault is given, where S122, S8, S9 
etc, shown in the first column of the table, are sensor 
labels. The rest of the columns in the table illustrate 
the association of attributes with the corresponding 
values for various sensors. The association of an at- 
tribute with a value of a sensor is called a feature. 
The columns two to four illustrate atomic features 
which involve only one attribute. The other three 
columns illustrate the compound features which are 
conjunctions of two atomic features. There are dif- 
ferent forms of attribute values in the Table. For ex- 
ample, sensor S122's attribute direction has a single- 
form of value POS; sensor pair S38/S41's attribute 
Rate-ratio has a range-form of value 4..5; sensor pair 
S9/S10's attribute Magnitude-ratio has a or-form of 
value 5V1. Similarly, in Table 5.5.a part of the dis- 
criminant description of MFV is given, which consists 
of features possessed by MFV's instances but not by 
any instances of other faults. 

VI. Conclusion 

We have developed an inductive machine learn- 
ing system, MLS, for the acquisition of knowledge 
about SSME faults. Given fault data from an en- 
gine simulator as input, MLS will generate charac- 
teristic, discriminant and aggregational descriptions 
about each engine fault. MLS also generates a con- 
cept hierarchy which groups related faults into clus- 
ters. The descriptions about each cluster are higher 
level descriptions of fault groups. The output from 
MLS can be used for assisting engineers in analyzing 
engine tests and for engine fault diagnosis. 

We have tested MLS with 61 fault instances from 
the SSME Simulator. Five valve blockage faults are 
included in those instances. MLS can correctly clas- 
sify the faults with a high rate of success. Human 
oriented descriptions about faults and fault groups 
are generated. 

Domain knowledge plays an important role in 
MLS. More knowledge enables the learning system 

to become more efficient and more accurate. For its 
future development, more domain knowledge about 
the engine needs to be added. Further research ef- 
forts are also needed to combine AI techniques with 
traditional statistical data analysis techniques. 


This work was performed within the Center for 
Advanced Space Propulsion (CASP) and was sup- 
ported in part by NASA Grant NAGW-1195, NAG- 
1-513 and Rocketdyne Contract No. R04QBZ90- 
032709. The Center for Advanced Space Propulsion is 
part of The University of Tennessee-Calspan Center 
for Aerospace Research, a not-for-profit organization 
located at UTSI. The authors would like to thank 
A.M. Norman for his assistance during the course of 
this effort. 

OPV Ho . 9 

CCV Mo. 5 

KFV Ho. 9 

o.neo — 

/ 'N_ 

Figure 5.1 Sensor Behavior. 


(d-cui«<.i£ {(direction f?eompi; o pos) 
(direction C7coep21 « .posl*) 
then Add ( (increAse-togethec (Tcompl ?cottp2) J ) ) 

(d-ruie5 if ((direction Ucompl) » neg;) 
(direction {?coatp2) » flea/)) 
then Add { (decreese-co^ether (icofnpl 2coiap2]))) 

{d-rule< if ({direction <?ca«pl) - poe) 
(direction <?conp2} - n«Q)l 
th«n Add {( oppoalte-trend (?co(npl 2comp2H)) 

(d-rule7 if (f't«rt-cljiw (Icompl) • Tstl) 
(stArt-tine <?eaetp2) - ?*t21 
(- ?etl !«t2ll 
then «dd { (chAnqe-concucrentiy (?conpl ?eonp2IH) 

Figure 5.2 Deductive Rules for Generating 
Component-Component Relationships. 

(d-cui«l if ( (etArt-tiJne (?coBp) - Jx) 
(end-time (?cotttp) - ?y|) 
then Add ( (duretion acoitp) - (- jy jx( ) J » 

(d-rulel2 If ((»c*rt-tlme (?»i» - ?vl» 
(AtArt-time (2l2) ■ Iv2) 
«- Ivi (V2II 
then Add {{ActlvAted-beforc (j»i J«2>)j) 

(p«r< At (faAqnleudA (a23> - ?ml) 
tluanitudA' (•21) - ?»2I I 
th«n Add 

UmA^nitudA-CAClo (i23/>27) - (/ }«a ?»2)l)) 

Figure 5.4 Hierarchy of Induction with Differentiat- 
ing Attribute Worth and Component Relevancy. 

(d-rulell If ((atACt-ttn (leoopil • Tvl I 
Action* ((Add-oinAAX 'first]}) 

(ddO if ((direction (»« - Irll 

(direction (a3«) - Tr2>) 
then Add ({And (direction (»9) - Jrl) 

(dlcAction (sjl) - ?r2))H 

(rrO if ((rete (»«) • Iril 

(rAt. (all) - ?r2l) 
then Add (IrAtA-tAtU (al/>2ll - (/ Irl lr2)II) 

Figure 5.3 Deductive Rules in ETID. 













starting time of an event 

ending time of an event 

average changing rate of an event 

indicate changing trend of an event 

the absolute amount of change of an event 

the average ..value of an event 

indicate whether ..the average value of an 

event is above, same of or below the 

normal value 
indicate the general pattern of the curve 
the value of the sensor after it became 

indicate whether the stable value of a 

sensor is above, same of or below the 

normal value 

the number of events for a sensor 

the difference between the START-TIME 


Table 5.1 Attributes Directly Extracted from Raw 






































Table 5.2 Sensor-Fault Relevancies. 





& value-to-normal 

& direction 

& stable-level- 




lower, less 




lower, less 




6. 1 


1. s 

Table 5.5 Discriminant Description of MFV. 














both events are rising together 
both events are (ailing together 
two events have the opposite trend 
two events have the same starting 

two events have the same starting 
time and ending time 
the event has the earliest starting 

one event activated before the other 
the event has the largest duration 
the ratio of two events' rate 
the ratio of two events' magnitude 
the ratio of two events' AVG-VALUE 
the ratio of two events' stable level 
the difference of two events' starting 

the difference of two events' ending 

Table 5.3 Derived Attributes. 





4 value-to-normal 

& direction 

& stable-level- 




higher, greater 






higher, greater 

S9, S10 



5 VI, 5 

Table 5.4 Characteristic Description of MFV.